Humanoid Robot Dodges 19/20 Thrown Balls Using Onboard Sensors

ChongZzZhang · x · 2026-08-01

The PAC-MAN project successfully trained a humanoid robot to play dodgeball. The system relies solely on the robot's onboard depth vision and proprioception, utilizing Control Barrier Function (CBF) guidance during training.

In zero-shot real-world tests on a Unitree G1 robot, it successfully dodged 19 out of 20 hand-thrown balls without a single fall.

Original post →

More from Embodied

Embodied channel →